Your team is using AI for a huge chunk of their work. Your CFO wants to know what that bought them. Do you have an answer?
If you're not sure, you're not alone — and that's exactly the problem.
The Numbers Tell Two Very Different Stories
Here's the 2026 developer AI landscape in two stats that have nothing to do with each other, and yet everything to do with each other:
84% of developers use AI tools at least weekly. (Stack Overflow.)
5% of enterprises see substantial ROI from AI at scale. (Not a typo.)
Those two numbers are living in completely different meetings. The first one is in your standup, your PR reviews, your team retros. The second is in your CFO's Q2 budget review, right next to a Forrester note that up to 25% of planned AI spend is being deferred to 2027 by organizations that couldn't prove returns in H1.
Your team has never been more bought-in on AI. Finance has never been more skeptical. That's not a contradiction — it's a measurement problem. And as an engineering leader, it's yours to solve.
The Vanity Metric Trap
When teams get asked to justify their AI spend, most of them reach for the wrong numbers. Prompts sent. Suggestions accepted. Copilot acceptance rate. Hours "saved" by survey.
These are vibes with a spreadsheet attached.
The trouble with "suggestions accepted" as a metric is that it tells you about usage, not value. A developer accepting every Copilot suggestion isn't a productivity win — it's a code review bill deferred to next quarter. An AI-generated PR that clears a linter but ships a subtle race condition hasn't saved you anything. You've just moved the cost downstream.
CFOs don't care how many times your team talked to an AI. They care whether AI changed the economics of shipping software. That requires a different kind of measurement.
What ROI Actually Looks Like in Engineering
The metrics that translate AI productivity into business outcomes are ones you probably already track — they just need to be viewed through the AI lens:
- Cycle time. Time from commit to deploy. If AI is compressing this, you should see it in your pipeline data. If you're not seeing it, either the AI isn't helping as much as your team thinks, or you have bottlenecks that have nothing to do with coding speed.
- PR throughput and size. Are your engineers shipping more frequently? Are PRs scoped more tightly? Smaller, more frequent PRs with AI assistance is a signal. Larger PRs that "just happened to use Copilot" is not.
- Defect rate and rework. The dirty secret of AI-assisted coding is that faster code isn't always better code. Track defect density per feature. If AI is helping your team ship fast but break things, your ROI calculation needs a quality term in it.
- Time-to-first-review. How long do PRs sit before they get eyes on them? AI can generate code faster than humans can review it — and if review is your bottleneck, AI-assisted coding is just building up a queue.
- On-call incidents per release. Blunt, but honest. If AI assistance is improving engineering quality, your on-call load should eventually reflect it. None of these require a new tool. They require you to draw a line in the data before and after rolling out AI tooling and actually look at what moved.
The Conversation You're Avoiding
Most engineering leaders haven't had a direct conversation with finance about what AI success looks like in terms finance understands. They've said "we're using AI, we're more productive" and hoped that would be enough.
It was enough in 2024. It's not enough now.
Forrester's warning about deferred spend isn't abstract. It means somewhere, a CFO looked at the AI line item and said show me. The engineering leaders who get ahead of this are the ones who go into that meeting with data, a before/after story, and a specific number — not a feeling.
"Our cycle time dropped 18% since Q3. PR throughput is up. Incidents per release are flat despite 30% more shipping velocity. Here's the trend." That's a conversation. "The team really likes Cursor" is not.
If Your AI Metrics Feel Like Vibes Right Now
Start small. Pick one team. Define three outcome metrics before you run any AI tooling experiment. Measure them for a quarter before, a quarter after. Report the delta to your skip-level with no editorial gloss — just the numbers and what you think they mean.
That's not a rigorous study. But it's the beginning of a language that finance and engineering can actually share.
The teams using AI to go faster are already everywhere. The teams that can prove they're going faster — and connect that to money — are the ones who'll have budget in 2027 to keep doing it.
Sources: Stack Overflow 2025 Developer Survey · Masterofcode — Enterprise AI ROI · Forrester via CIO.com · LinearB 2026 Engineering Benchmarks